[Next] [Previous] [Top] [Contents]
Towards a Descriptive Model of Agent Strategy Search - Bruce Edmonds
What makes Sonnemans' experiment particularly interesting is that it allowed the subjects reasonable scope within which to express their strategies. Even though the scope was still fairly restrictive (as compared with natural language), the subjects specified a range of different kinds of strategies. Clearly, the subjects are learning the form of their strategy as well as its parameterisation. This fits into the formal framework for learning proposed in [9]. He also has collect some statistics about their behaviour at different stages of the experiment (Table 2). However these are not sufficient to constrain the possible learning mechanisms employed by the subjects (it was, of course, not designed with this in mind*1). Thus I have had to look elsewhere for additional constraints - in this case, cognitive science.
I am not suggesting that economics should (or could) be reduced to cognitive science, just that we should accept constraints about the behaviour of our units from it - just as chemistry ensures that its models are consistent with what physics tells them about the behaviour of atoms. Nor am I suggesting that we should invade the field of cognitive science - our job is to ensure that the behaviour of our agents matches that of real economic actors so that we can safely explore what happens when they interact monetarily. The job of investigating the purely internal cognitive processes is best left to cognitive scientists. However, just as good chemists must know enough about the physics of atoms in order to ensure that their chemical models are consistent with the laws of physics so must we inform ourselves sufficiently about cognitive science in order to ensure that our agents are compatible with what is known about the workings of humans.
My source is `Induction' by Holland, Holyoak, Nisbett and Thagard [5]. This is a synthesis of the relevant cognitive science and associated philosophy on how induction does and can occur, written from a computational point of view. This attempts to cover all the different aspects of induction (e.g. induction in science), so I will only draw on the basic framework that it suggests. In particular, I have abstracted the following aspects:
- That induction works on a population of `rule-based mental models';
- That there are operations that produce new models from the existing ones;
- That there is some `credit assignment' mechanism that allows the hypotheses to be evaluated in response to their success.
Although [5] uses the paradigm of the `classifier system' as a concrete framework to explore and illustrate the possibilities, I will use the Genetic Programming (GP) paradigm [7]*2. This provides a flexible framework to implemented this picture of induction in an artificial agent. It is more appropriate for this task because the form of the strategies that Sonnemans reported can be directly implemented using GP-based techniques. The GP paradigm involves a collection of expressions that conform to a formal grammar, which are evolved by the operations of combination, variation, propagation and selection, in response to the tasks it is presented with. The original GP algorithm was designed for efficient machine induction, so there is no reason, a priori, to suppose that it is a good picture of human induction. However the paradigm is much wider, providing a framework within which there are many possible algorithms. This flexibility enables us to search for a descriptive fit with the observed results. The use of this paradigm is examined in more detail elsewhere [3].
Towards a Descriptive Model of Agent Strategy Search - Bruce Edmonds - 06 SEP 99
[Next] [Previous] [Top] [Contents]
Generated with CERN WebMaker